[PATCH 2/9] ARM: SPMP8000: Add machine base files

[Jean-Christophe PLAGNIOL-VILLARD]

On 22:59 Sun 16 Oct     , Arnd Bergmann wrote:
> On Sunday 16 October 2011, Zoltan Devai wrote:
> > 2011/10/11 Arnd Bergmann <arnd at arndb.de>:
> > > On Sunday 09 October 2011, Zoltan Devai wrote:
> > >> +
> > >> +/* Static mappings:
> > >> + * SCU: timer needs clk support which is inited in init_early
> > >> + * UART: needed for earlyprintk
> > >> + */
> > >> +static struct map_desc io_desc[] __initdata = {
> > >> +     {
> > >> +             .virtual        = IO_ADDRESS(SPMP8000_SCU_A_BASE),
> > >> +             .pfn            = __phys_to_pfn(SPMP8000_SCU_A_BASE),
> > >> +             .length         = SPMP8000_SCU_A_SIZE,
> > >> +             .type           = MT_DEVICE,
> > >> +     }, {
> > >> +             .virtual        = IO_ADDRESS(SPMP8000_SCU_B_BASE),
> > >> +             .pfn            = __phys_to_pfn(SPMP8000_SCU_B_BASE),
> > >> +             .length         = SPMP8000_SCU_B_SIZE,
> > >> +             .type           = MT_DEVICE,
> > >> +     }, {
> > >> +             .virtual        = IO_ADDRESS(SPMP8000_SCU_C_BASE),
> > >> +             .pfn            = __phys_to_pfn(SPMP8000_SCU_C_BASE),
> > >> +             .length         = SPMP8000_SCU_C_SIZE,
> > >> +             .type           = MT_DEVICE,
> > >> +     },
> > >
> > > With Nicolas Pitre's rework of the MMIO space handling, I think it
> > > would be nice to just use a single area that spans all of the devices
> > > so you can do an optimized ioremap and huge TLBs. Unfortunately
> > > that series won't make it into 3.2, but if you just set up a large
> > > area now, it will automatically work.
> > What do you mean by 'all of the devices' ?
> > These three, or all MMIO peripherals ?
> > These are at addresses 0x90005000, 0x92005000 and 0x93007000.
> > So either a 50 or 256 MB area, but both seem like a waste of
> > virtual memory space.
> 
> I was thinking of mapping everything that your IO_ADDRESS macro covers,
> which should be something like 240 MB starting at 0x90000000/0xf0000000.
> On most ARM9 platforms, virtual memory is not really contraint at all
> because there is not all that much physical memory. Also, you currently
> reserve the entire 0xf0000000 segment for MMIO ranges, so I would
> expect that you can keep doing that after the cleanup.
> 
> > >> +#ifdef CONFIG_DEBUG_LL
> > >> +     {
> > >> +             .virtual        = IO_ADDRESS(SPMP8000_UARTC0_BASE),
> > >> +             .pfn            = __phys_to_pfn(SPMP8000_UARTC0_BASE),
> > >> +             .length         = SPMP8000_UARTC0_SIZE,
> > >> +             .type           = MT_DEVICE,
> > >> +     },
> > >> +#endif
> > >> +};
> > >>
> > > And it would be nice to move this into a separate file that handles the
> > > early debug code, as prima2 does.
> > Prima2 doesn't have other static mappings. In my case, that would
> > mean a separate map_desc table in lluart.c and an
> > #ifdef CONFIG_DEBUG_LL in the map_io call.
> > Is it worth it ?
> 
> Yes, it's worth it. Don't do the #ifdef in the map_io call though, do
> it in the header file that declares the function prototype.
> 
> > BTW, the static mappings are only needed to be able to set up the
> > clk stuff in init_early. That in turn is only needed for the timer code to
> > determinate its input clock.
> > If I would hard-code the timer clock rate in its driver (the bootloader is
> > very unlikely to change), then I could init the clk driver from
> > machine_init and get rid of these mappings.
> > How about that ?
> 
> Maybe you can just pass the clock rate in the device tree then. If the boot
> loader changes it, it can still pass the correct rate and you don't
> need the early mappings for functional reasons any more.
please don't update the DT from the bootloader, I'll recommand you to limit
your bootloader to just update the strict minimun (same as ATAG) and touch
noting else

Best Regards,
J.